QUANTIFICATION OF MUSCLE FATIGUE WITH WAVELET

ANALYSIS BASED ON EMG DURING MYOELECTRICAL

STIMULATION

M. Yochum

, T. Bakir

, R. Lepers

and S. Binczak

LE2I CNRS UMR 5158, Universit

e de Bourgogne, 9 avenue Alain Savary, BP47870 21078 Dijon cedex, France

INSERM U887, Universit

e de Bourgogne, BP 27877 21078 Dijon cedex, France

Keywords:

Fatigue, Electromyogram, Wavelet, Electrical stimulation.

Abstract:

We propose a device dedicated to real time analysis of electromyograms (EMG) under myoelectrical stimu-

lation (ES). The muscular fatigue analysis, which is obtained by the use of a dedicated analog circuit and a

processing part, is the main purpose of this study. The description of a hardware device which incorporates

an electro-stimulator and an electromyogram ampliﬁer combined to a computer is detailed. Then, we present

a muscular fatigue analysis part based on wavelet decomposition in order to extract a fatigue index, which is

confronted with synthetic and experimental data. We conclude that the CWT index applies well to M waves.

The noise sensitivity is investigated.

1 INTRODUCTION

Devices using the myoelectrical stimulation (ES)

and the analysis of electromyograms (EMG) become

more sophisticated and their applications continue to

grow. It is the case especially for muscular reha-

bilitation systems (Chilibeck et al., 1999) for peo-

ple who have had nerves trauma like paraplegics or

hemiplegics and for people who have had a tempo-

rary immobilization of one or more limbs leading up

to muscle atrophies. Other applications allow FES as-

sisted gait (Matjai et al., 2003; Iwami et al., 2009;

Hunt and Stone, 2004), equally in medicine with the

intention of search myasthenia or muscular degener-

ations (Stein and Mushahwar, 2005). The ES retards

and even inverses the muscular atrophy. The ﬁrst goal

is to increase the muscle mass by artiﬁcial contrac-

tions of muscles. In this way, muscles react as dur-

ing a real exercise and create more muscular tissues

(Erika Scremin et al., 1999).

Electrical stimulators have considerably changed

since their discovery (Dolhem, 2008) and some fol-

low a predeﬁned program with current controlled in-

jection (Velloso and Souza, 2007). EMGs allow ob-

taining some informations about the muscle contrac-

tions. These informations can give the state of a mus-

cle or its changes over time. Recent works show the

use of those indexes in order to manage a dedicated

electrical stimulation (Yeom and Chang, 2010; Sink-

jaer et al., 2003). The fatigue analysis during a mus-

cular effort can be essential. When a patient does

not feel the fatigue, it can be transcribed by an in-

dex resulting from the evaluation of EMG (Mizrahi

et al., 2002). During ES, a typical electrical response

appears on EMGs: the M Waves (Chesler and Dur-

fee, 1997). This is usually a biphasic wave. It is

proved that those wave shapes are changing during

an electrical stimulation (Thomas, 1997). It is these

changes which are used to determine the muscular fa-

tigue (Knaﬂitz and Deluca, 1990).

Various indices of muscle fatigue are available

with different types of signals. The most common

are performed on EMGs which are acquired dur-

ing voluntary or electrically stimulated contractions

(Mizmhi, 1997). Others indexes use the muscle

strength developed (Tepavac and Schwirtlich, 1997).

In our case, for EMGs, one can cite the Peak to Peak

(PTP), Root Mean Square (RMS), Mean Frequency

mean

) and Median Frequency (F

med

) or also the

power spectral density and the zero crossing (Mizmhi,

1997; Tepavac and Schwirtlich, 1997; Chesler and

Durfee, 1997).

Some recent articles apply wavelet treatments for

fatigue analysis. It is, mostly, obtained by discrete

wavelet transform (DWT) where common wavelets

are used (Kumar et al., 2004; Vukova et al., 2008).

Yochum M., Bakir T., Lepers R. and Binczak S..

QUANTIFICATION OF MUSCLE FATIGUE WITH WAVELET ANALYSIS BASED ON EMG DURING MYOELECTRICAL STIMULATION.

DOI: 10.5220/0003778600530058

In Proceedings of the International Conference on Biomedical Electronics and Devices (BIODEVICES-2012), pages 53-58

ISBN: 978-989-8425-91-1

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

In this context, we introduce in section 2 a new

electro stimulator with EMG feedback and a software

which work in real time. Many fatigue treatments

were implemented, some from literature and a new

one, based on continuous wavelet transform (CWT),

which is detailed in section 3. Tests have been made

on synthetic and experimental EMG, which are pre-

sented in section 4 while in section 5, the inﬂuence of

noise is also investigated because EMGs are very little

signals easily corrupted by the noise. Finally section

6 discusses the presented results on real experimental

data and section 7 concludes this article.

2 MATERIAL

We introduce a new electro stimulator allowing the

electrical stimulation of a muscle and the fatigue anal-

ysis applied to the EMGs feedback in real-time. The

device is composed of a hardware part dedicated to

deliver stimulations and EMGs ampliﬁcations. A

software part allows the control of the stimulation and

computes the fatigue index. A NIDaq module con-

nects these two parts making the system processing

in real-time. Hardware is composed of two parts.

The ﬁrst one is a stimulation board whose circuit is

represented on Figure 1. In order to have the same

stimulation for all muscles, we decided to use con-

troled injection currents. The circuit is inspired by the

work of Han-Chang Wu & al (Wu et al., 2002). The

board works symmetrically, the top part is for positive

phases and the bottom part for negative phases. The

software generates stimulation pulses from −10 V to

10 V . The OPAs copy the voltage V

onto V

opa

with a

null input current. Then, pulse voltages are converted

in pulse currents thanks to R1 and Q1 in order to ob-

tain Ic such as Ic =

opa

. This current can be

maintained thanks to the high tension on V DD and

−V DD. Ic is then copied into the stimulation elec-

trodes by the Wilson current mirror (Q

, Q

, R

and R

Figure 2 shows the EMG board circuit which is

the second hardware part. We use surface electrodes

because it is a non invasive and painless equipment,

although the electrodes positions are crucial (Rutkove

et al., 2005). Indeed, EMG signals vary according to

their positions on the skin (Rutkove et al., 2005). Two

electrodes are on the muscle (E

and E

) and another

laid on a bony point (E

re f

) acting as a reference volt-

age. This board makes the difference between two

voltages of the muscle with a large ampliﬁcation and

reference removal. The two instrumentation ampli-

ﬁers (INA) U

and U

are used to erase the reference

voltage of the body for E

and E

Figure 1: Stimulation diagram.

= G

·(E

−E

re f

), with G

, (1)

where i = {1,2} indicates which electrode is con-

cerned and RP

is in KΩ. In many EMG ampliﬁers,

just one INA is used to obtain the difference between

the two muscular electrodes (Mesin et al., 2009). In

our case, we have access to the muscular activity of

each electrode. Then, the difference between them is

obtained by the same method (eq. 1), consequently

EMG

= G

·(V

−V

). Note that V

EMG

can be ﬁl-

tered with a pass band ﬁlter.

The software is developed with Labview. It al-

lows to choose many stimulation parameters in or-

der to have different stimulation pulse shapes and se-

quences. The current amplitude can vary from 0 mA

to 100 mA, the duration of pulses can be set from

500 µs to 2000 µs, the frequency of pulse train from

10 Hz to 100 Hz. The shape of pulses have been

chosen among the most common ones in the litera-

ture (Watanabe et al., 2000; Karu et al., 2002) such as

Monophasic ones, Biphasic ones, Dual Biphasic ones,

Asymetric Biphasic ones and Doublet Nlet ones. The

stimulation and rest duration and the kind of fatigue

treatment are also adjustable. Many graphics have

been added in order to visualize the electrical activity

such as the representation of one period of stimula-

tion pulses, the ongoing EMG signal in real time, the

reference M wave and fatigue analysis results.

BIODEVICES 2012 - International Conference on Biomedical Electronics and Devices

Figure 2: EMG diagram.

3 FATIGUE DETERMINATION

The EMG board leads to obtain the muscular electri-

cal activity but also the stimulation artifacts (Fig 3.a).

The elimination of these artifacts are delicate because

M waves are in the same ranges of frequency. The

method of artifacts removal which has been chosen is

a two-stage peak detection algorithm (O’Keeffe et al.,

2001). A signal with this artifact removal is given as

example in Fig 3.b. The ﬁrst M wave is used to con-

struct a wavelet pattern from which the dilation of the

folowing M waves is deduced by using the continuous

wavelet transform (CWT), that is:

a,b

s(t)

√



t −b



dt, (2)

8.3 8.35 8.4 8.45 8.5

−10

Original EMG Signal

8.3 8.35 8.4 8.45 8.5

−1

EMG Signal without Artifacts

8.3 8.35 8.4 8.45 8.5

100

200

Time (s)

Scale a

Local Maxima

Figure 3: Example of artifact removal. a. EMG signal from

EMG board b. EMG without artifacts c. Local maxima cir-

cles represent the values of CWT coefﬁcients in top vision

like topographical view. Big circles are small values and

small circles are high values.

with s(t) the signal, (i.e. the EMG signals), ψ the used

wavelet (in our case, it is the ﬁrst M wave) and a

the applied scale factor. For each received M wave

during the stimulation, a local maxima algorithm is

used on the results of CWT, leading to ﬁnd the best

match between the scale factor and the temporal ex-

pansion of M waves. Figure 3.c shows some levels

of CWT coefﬁcients, the smallest circles correspond-

ing to the highest values. Those indexes represent the

expansion undergone by the M wave reference dur-

ing the stimulation. The scales from local maxima

are used as fatigue indexes. The indexes of fatigue

from literature move downward zero over time, con-

trary to CWT indexes which go up from 1. In order

to keep this tendency, the inverse of scale parameters

have been taken as I

CWT

= 1/a. Resulting curves are

displayed in Figure 7. In addition to our fatigue index

based on the CWT, other indexes were implemented

in the software. Those treatments are directly in-

spired from literature (Thomas, 1997; Mizmhi, 1997;

Chesler and Durfee, 1997; Mizrahi et al., 1997). Two

indexes are based on magnitude, Peak To Peak (PTP)

and Root Mean Square (RMS) and two others are fre-

quency based, Mean Frequency (F

mean

) and Median

Frequency (F

med

4 TEST WITH SYNTHETIC

SIGNALS

Synthetic EMG signals have been realized to check

the CWT fatigue algorithm by using a M wave refer-

ence recorded experimentally. Then, thanks to those

reference M waves, synthetic EMGs are created. Fig-

ure 4 illustrates the method for the experimental M

wave (Fig 4.a). The M wave reference is temporally

increasingly dilated by a factor from 1 to 3 (Fig 4.b).

Then, each resulting M wave is juxtaposed (Fig 4.c)

and weighted by a decreasing exponential envelope to

give a synthetic realistic EMG (Fig 4.d).

The CWT algorithm is applied on the synthetic

EMGs. Figure 5 shows the results for experimental

based EMG signal. To check the efﬁciency of this al-

gorithm, a mean-square error measure is performed.

This error is computed as:

mean

∑

i=1

−IR

×100, (3)

where ER is means of error ratio, FR are found results

and IR are ideal results. n represents the number of

analysed M wave. Table 1 at the line ”without noise”

displays this error ratio, which is equal to 1.04% for

the synthetic EMG based on an experimental M wave.

QUANTIFICATION OF MUSCLE FATIGUE WITH WAVELET ANALYSIS BASED ON EMG DURING

MYOELECTRICAL STIMULATION

The error rate indicates that the CWT based index ap-

plies quite well to experimental M waves.

5 NOISE SENSITIVITY

EMGs are very little signals that can engender a sig-

niﬁcant noise to signal ratio (NSR). Therefore, it may

be useful to study the inﬂuence of noise on the pro-

cess. A standard uniform noise distribution on the

open interval (-1,1) is used. Then it is multiplied by

a factor deﬁned on a certain percentage of maximum

voltage of EMG and the EMG signal and noise are

added. Figure 6 shows the CWT fatigue indexes on

a experimental based EMG signal under noise (the

magnitude of noise was 10 % of the V

EMG

maxima).

The dot line corresponds to the ideal case, the dashed

line to a noisy EMG signal and the continuous line to

a pure EMG one. Obviously, the results with noisy

EMG are worse than pure EMG.

With the intention of reducing the noise impact

on the CWT based index, ﬁve ﬁlters have been im-

plemented. Those ﬁlters can be separated into two

groups; the ﬁlters of the ﬁrst group are applied to

EMG signals. In this case, the ﬁltering which is ap-

plied before the fatigue extraction can be a simple low

pass 1D Butterworth one or a 1D wavelet one which

uses the discrete stationary wavelet transform (SWT).

The second group corresponds to some ﬁltering ap-

plied to the CWT coefﬁcients. Then, ﬁltering is made

during the fatigue algorithm computing. Three cases

have been investigated; the ﬁrst one is an imagery ﬁl-

tering issue, the second one is a 1D Butterworth ﬁlter

applied to each line of the 2D matrix C

a,b

and the last

one is a 2D wavelet ﬁltering based on SWT. Filters are

0 50 100 150 200 250 300 350 400 450 500

−1

Experimental M−wave

0 200 400 600 800 1000 1200 1400

−1

Experimental M−wave extended

0 1000 2000 3000 4000 5000 6000 7000 8000

−1

Experimental Artificial EMG

0 1000 2000 3000 4000 5000 6000 7000 8000

−1

Experimental Artificial EMG with exponential decreasing

Figure 4: a. M wave from experiment b. M wave extended

in time with a factor from 1 to 3 c. artiﬁcial EMG signal

created with experimental extended M waves by juxtaposi-

tion d. loss in amplitude by an decreasing exponential to

give ﬁnal artiﬁcial EMG signal.

0 50 100 150

0.4

0.5

0.6

0.7

0.8

0.9

N° M−wave

Scale Parameter (1/a)

Ideal results

found results

Figure 5: Results of fatigue detection by CWT with syn-

thetic EMG (gray + line) and ideal result (black line).

0 50 100 150

0.4

0.5

0.6

0.7

0.8

0.9

N° M−wave

Scale Parameter (1/a)

Ideal Results

Noisy Result

Pure Result

Figure 6: Application of the CWT fatigue index on an ex-

perimental based EMG signal. Dot line are the ideal results,

dash line are the results for the noisy EMG (the magnitude

of noise was 10 % of the V

EMG

maxima) and continuous

line for the EMG signal.

applied to the synthetic EMG signals. In order to have

a clear representation of ﬁlters efﬁciency, averages of

errors have been made by using eq. 3. Those errors

are listed in Table 1. For both Butterworth ﬁlters, the

errors are identical all the time, but the execution du-

ration is really longer for the 2D ﬁltering because it is

applied to the CWT coefﬁcients. Therefore, the num-

ber of point that need to be treated is much more im-

portant. The best ﬁlter for the synthetic EMG signals

is the imagery inspired one (a circular averaging ﬁl-

ter) with a decrease from 6.69% of error for the noisy

signal to 2.48%.

The same test on the inﬂuence of noise has also

been performed for the others fatigue indexes from

the literature. As 2D ﬁltering cannot be applied to

those signals because no CWT has been made, only

two ﬁlters have been performed; the 1D Butterworth

and the 1D SWT ﬁlter. The errors between the ideal

results and three signals (noisy synthetic EMG, ﬁl-

tered with Butterworth and SWT) have been made

and those errors are shown in Table 2. The differ-

BIODEVICES 2012 - International Conference on Biomedical Electronics and Devices

ence of error between the results of a unﬁltered signal

and a ﬁltered signal is very signiﬁcative. The error is

more signiﬁcative for the SWT ﬁltering. It is there-

fore necessary to choose a SWT ﬁlter rather than a

classic one because wavelet ﬁlter are more efﬁcient

on a noisy EMG.

Tables 1 and 2 show that CWT fatigue indexes are

less noise dependant than the literature indexes with

or without denoising.

Table 1: Square-Mean Error for the CWT fatigue index (%).

Kind Experimental

Filtre Based Error

Without Noise 1.0395

With Noise 6.6945

Filter 1D Butterworth 2.6042

Filter 1D SWT 2.8618

Filter 2D Image 2.4826

Filter 2D Butterworth 2.6042

Filter 2D SWT 7.8229

Table 2: Square-Mean Error for Indices in Litera-

ture(%),PTP: peak to peak, RMS: root mean square, Fmean:

mean frequency and Fmed: median frequency.

Experimental Based EMG Signal

Fatigue Noisy Butterworth SWT

index Signal Filtered Filtered

PTP 34.29 5.36 3.37

RMS 68.61 12.67 3.62

Fmean 204.48 38.60 11.61

Fmed 173.18 23.90 3.82

6 VALIDATION

The CWT fatigue processing has been applied to

EMG signals obtained under experimental exercice

which has been made on the right biceps. The stim-

ulation pulses were biphasic and symetric and the

pulses duration was 1000 µs. The pulse train was

50 Hz and the injected current was 60 mA. The ar-

tifacts have been removed and the ﬁrst M wave found

has been used as a reference M wave. The results of

CWT treatment are shown in Figure 7 for two dis-

tinct subjects. The indexes go toward more and more

weak values. The dilation of M waves becomes in-

creasingly important over the stimulation. As the ex-

pansion of the M waves is an indicator of fatigue, this

curve shows that the muscular fatigue increases dur-

ing ES. The fatigue indexes based on CWT are in-

deed an indicator of fatigue usable on EMG signals

during ES. On the bottom result (Fig 7), the fatigue

0 1 2 3 4 5 6 7 8 9

0.7

0.8

0.9

Scale Parameter (1/a)

0 1 2 3 4 5 6 7 8 9 10

0.7

0.8

0.9

Time (s)

Scale Parameter (1/a)

Figure 7: Results of the fatigue detection thanks to our

CWT treatment for two distinct subject. On the bottom plot,

we can see a plateau time at the beginning of the stimula-

tion.

indexes rest to its maximum level during beginning of

the stimulation, corresponding to a plateau state.

7 CONCLUSIONS

A presentation of a ES device with the analyze of fa-

tigue thanks to an EMG feedback and a CWT treat-

ment has been given. It has been shown that wavelets

can provide a reliable fatigue index by using the drift

of M wave elongation during a ES. A study of noise

sensitivity was made and showed that imaging ﬁlters

(a circular averaging ﬁlter) apply rather well to CWT

fatigue indexes. It would be interesting to conﬁrmed

these primilary results on an increased number of sub-

jects and muscles. Subjects may be healthy or with a

disease in order to compare the variation of indexes

between them.

REFERENCES

Chesler, N. and Durfee, W. (1997). Surface EMG as a fa-

tigue indicator during FES-induced isometric muscle

contractions. Journal of Electromyography and Kine-

siology, 7(1):27–37.

Chilibeck, P., Bell, G., Jeon, J., Weiss, C., Murdoch, G.,

MacLean, I., Ryan, E., and Burnham, R. (1999).

Functional electrical stimulation exercise increases

GLUT-1 and GLUT-4 in paralyzed skeletal muscle 1.

Metabolism, 48(11):1409–1413.

Dolhem, R. (2008). Histoire de l’

electrostimulation en

edecine et en r

education. Annales de R

eadaptation

et de M

edecine Physique, 51(6):427–431.

Erika Scremin, A., Kurta, L., Gentili, A., Wiseman, B.,

Perell, K., Kunkel, C., and Scremin, O. (1999). In-

creasing muscle mass in spinal cord injured persons

with a functional electrical stimulation exercise pro-

gram 1. Archives of physical medicine and rehabilita-

tion, 80(12):1531–1536.

QUANTIFICATION OF MUSCLE FATIGUE WITH WAVELET ANALYSIS BASED ON EMG DURING

MYOELECTRICAL STIMULATION

Hunt, K. and Stone, A. (2004). Control strategies for inte-

gration of electric motor assist and functional electri-

cal stimulation in paraplegic cycling: utility for exer-

cise testing and mobile cycling. IEEE Transactions

on Neural Systems and Rehabilitation Engineering,

12(1):89–101.

Iwami, T., Sasaki, M., Miyawaki, K., Matsunaga, T., Shi-

mada, Y., and Obinata, G. (2009). Biomechani-

cal analysis for FES assisted swing-through gait. In

Micro-NanoMechatronics and Human Science, 2008.

MHS 2008. International Symposium on, pages 430–

434. IEEE.

Karu, Z., Durfee, W., and Barzilai, A. (2002). Reduc-

ing muscle fatigue in FES applications by stimulat-

ing with N-let pulse trains. IEEE Transactions on

Biomedical Engineering, 42(8):809–817.

Knaﬂitz, R. M. M. and Deluca, C. J. (1990). Myoelectric

manifestations of fatigue in voluntary and electrically

elicited contractions. the American Physiological So-

ciety 0161-7567/90.

Kumar, D., Pah, N., and Bradley, A. (2004). Wavelet analy-

sis of surface electromyography. IEEE Transactions

on Neural Systems and Rehabilitation Engineering,

11(4):400–406.

Matjai, Z., Hunt, K., Gollee, H., and Sinkjaer, T. (2003).

Control of posture with FES systems. Medical engi-

neering & physics, 25:51–62.

Mesin, L., Merletti, R., and Rainoldi, A. (2009). Surface

EMG: The issue of electrode location. Journal of

Electromyography and Kinesiology, 19(5):719–726.

Mizmhi, J. (1997). Fatigue in muscles activated by func-

tional electrical stimulation. Crit. Rev. Phys. Rehabil.

Med, 9(2):93–129.

Mizrahi, J., Levin, O., Aviram, A., Isakov, E., and Susak,

Z. (1997). Muscle fatigue in interrupted stimulation:

effect of partial recovery on force and EMG dynam-

ics. Journal of Electromyography and Kinesiology,

7(1):51–65.

Mizrahi, J., Levy, M., Ring, H., Isakov, E., and Liberson,

A. (2002). EMG as an indicator of fatigue in isomet-

rically FES-activated paralyzed muscles. IEEE Trans-

actions on Rehabilitation Engineering, 2:57–65.

O’Keeffe, D., Lyons, G., Donnelly, A., and Byrne, C.

(2001). Stimulus artifact removal using a software-

based two-stage peak detection algorithm. Journal of

neuroscience methods, 109(2):137–145.

Rutkove, S., Partida, R., Esper, G., Aaron, R., and Shiffman,

C. (2005). Electrode position and size in electrical

impedance myography. journal of the International

Federation of Clinical Neurophysiology, 116(2):290.

Sinkjaer, T., Haugland, M., Inmann, A., Hansen, M., and

Nielsen, K. (2003). Biopotentials as command and

feedback signals in functional electrical stimulation

systems. Medical engineering & physics, 25(1):29–

40.

Stein, R. and Mushahwar, V. (2005). Reanimating limbs

after injury or disease. Trends in neurosciences,

28(10):518–524.

Tepavac, D. and Schwirtlich, L. (1997). Detection and

prediction of FES-induced fatigue. Journal of Elec-

tromyography and Kinesiology, 7(1):39–50.

Thomas, C. (1997). Fatigue in human thenar muscles paral-

ysed by spinal cord injury. Journal of Electromyogra-

phy and Kinesiology, 7(1):15–26.

Velloso, J. and Souza, M. (2007). A Programmable System

of Functional Electrical Stimulation (FES). In Engi-

neering in Medicine and Biology Society, EMBS 2007,

pages 2234–2237. IEEE.

Vukova, T., Vydevska-Chichova, M., and Radicheva, N.

(2008). Fatigue-induced changes in muscle ﬁber ac-

tion potentials estimated by wavelet analysis. Journal

of electromyography and kinesiology, 18:397.

Watanabe, T., Miura, N., Hoshimiya, N., and Handa, Y.

(2000). The possibility of using m-waves related to

double pulses for evaluating muscle fatigue in fes con-

trol. Japanese Journal of Medical Electronics and Bi-

ological Engineering, 38(1):42–48.

Wu, H., Young, S., and Kuo, T. (2002). A versatile mul-

tichannel direct-synthesized electrical stimulator for

FES applications. IEEE Transactions on Instrumen-

tation and Measurement, 51(1):2–9.

Yeom, H. and Chang, Y. (2010). Autogenic EMG-

Controlled Functional Electrical Stimulation, for An-

kle Dorsiﬂexion Control. Journal of Neuroscience

Methods.

BIODEVICES 2012 - International Conference on Biomedical Electronics and Devices